Most data fusion methods assume all data sources to be conditionally independent. In most cases, this is a practical choice, rather than an assumption derived from the observed phenomena. Wifi-based localization uses wireless signal strength information from all access points in an environment to estimate a robot's locations, which has been used both in indoor and outdoor (urban) environments. In such areas, the number of access points often ranges from several tens to a few hundred access points, several of which are strongly correlated to each other. In such a case, assuming all data sources are independent results in extremely peaked likelihood functions (overconfident) that are not suitable for robot localization under a Bayesian approach. Previous work has shown experimentally that a rather conservative, complete dependence assumption holds better in practice, as it is more robust against sensor and mapping errors. However, such an assumption generates less informative posteriors than desired, lowering localization accuracy. In this work, rather than taking either of these extremes, we propose a data-based method that learns more balanced data fusion rules, which generate informative yet robust likelihood functions. Testing performed on both indoor and outdoor datasets show the feasibility of our method.